Everyone's one-line explanation of AI detection is the same: it measures perplexity and burstiness. I went to read what the two vendors most often cited for that actually publish today. Neither page says it, and the way each one does not say it is different.
Turnitin's is a written denial:
"Our model is not explicitly programmed to evaluate specific signals such as “burstiness,” “perplexity,” or other individual metrics sometimes referenced in public discussions."
GPTZero's is stranger, because GPTZero is where this framing went mainstream. Neither term is its invention. Perplexity has been a standard language-model measure for decades and burstiness as a text statistic long predates any of this. But "AI detectors measure perplexity and burstiness" is the sentence that spread outward from GPTZero, and today its own /technology page and its own /faq page contain neither word, not once.
Roughly 12,900 and 13,700 characters of rendered text. perplexit* zero, burstiness zero, on both. Positive controls on the same counter in the same pass, case-insensitive substring matching throughout: detect 51 and 24, model 31. An impossible control string returned zero, as it should. If you recount case-sensitively the controls will come out different; the two zeroes hold either way, since neither word appears in any casing.
Go count it yourself, it takes a minute. Your totals will not match mine exactly, because those pages carry dynamic content and I got slightly different character counts on different passes. The two zeroes do not move.
Careful with what that means, though. Dropping a word is not retracting it. GPTZero has not said the framing was wrong. It has stopped describing itself that way. Those are two different claims and only the second one is mine to make.
The terms, quickly, since half the arguments are definitional
Perplexity is how surprised a language model is by the next token. Feed it The experiment showed a significant and the distribution piles onto result, difference, effect. Low surprise. Put pineapple in that slot and it spikes. Average over a document and you have one scalar for "how predictable were these choices".
Burstiness is variance in sentence shape. Uniform lengths and structures score low, a two-word sentence next to a forty-word one scores high.
Both are real quantities, and both are computable by you in an afternoon with any open-weights model and a tokenizer. That is exactly why they became the standard story. They are the legible part of the problem, and the legible part of a problem is what gets repeated.
What the two pages describe instead
GPTZero's technology page is a list of components, not metrics: an end-to-end deep learning approach trained on web, education and LLM-generated text, a sentence-by-sentence classification model producing a probability and a confidence, and a "Paraphraser Shield" it says defends against paraphrasing and homoglyph attacks. In its research bullets it frames detection as a trinary classification problem, separating prediction confidence from the proportion of LLM text, and claims the first sentence-highlighting model built on HMMs.
Whether any of that is true is not something I can check from outside. But notice the shape: it is an architecture description, and there is no scalar in it you could optimise against.
Turnitin's page goes further into plumbing:
"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated."
Sentences extracted, chunked into overlapping windows, each window scored in [0, 1], scores pooled up into a document-level percentage. The overlap is the detail I would not have guessed. The unit being classified is not the sentence you wrote. A sentence sits in more than one window and inherits from all of them.
The trap on the Turnitin page
That denial sits a long way down, under the question Which model is Turnitin's AI detection based on? — an architecture question, answered with a paragraph about transformers. About eighteen thousand characters earlier, near the top of the same page, there is a question that asks the thing most people actually mean: What parameters or flags does Turnitin's model take into account when detecting AI writing?
It is answered like this:
"Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers."
Perplexity is a word-probability measure. So anyone reading top to bottom meets the word-probability answer first, and only much later runs into a sentence that looks like its opposite.
It doesn't contradict it, and the distinction is one every ML engineer already has: the denial is about architecture, not about information. No feature-extraction step that computes a perplexity scalar and hands it to a classifier. The next sentence completes it:
"Instead, it learns statistical patterns from our training data."
A model trained on differences in word-probability sequences will encode plenty that correlates with perplexity. It just has no named tensor you can point at.
So: quoting the denial on its own to argue that word probability is irrelevant here is a losing move. The word-probability sentence is the one a reader hits first, sitting under the question that literally asks what the model takes into account. The two quotes have to travel together or not at all.
The same page is also unusually candid about the cost of that design:
"As a result, its outputs are generated by many learned patterns working together rather than by a small set of transparent, human-readable rules. For that reason, individual predictions may not always be explainable in simple feature-by-feature terms."
That is a vendor writing "our model is not interpretable" into its own customer-facing documentation, which I did not expect to find.
What to do with this
One thing, and I would rather name one real use than manufacture five:
Before you read anything about what a percentage means, establish which system produced it, then go and read what that system currently says about itself. Both of the pages above have moved. The popular explanation did not move with them, because an explanation that is easy to repeat outlives the thing it was explaining.
And treat every sentence I quoted as perishable, including the two zeroes. Capture it with the URL visible and a date attached. Otherwise you end up doing what I did earlier today, which was going back for saved quotes on a different vendor's help centre and finding seven of them gone.
If what you are holding is an exported report rather than a theory, HumanPen imports a Turnitin or iThenticate AI report, matches the flagged passages back to your document, and rewrites only the ones you confirm. The rest of the file stays untouched. Eligible passages can be re-run at no charge. I work on it: humanpen.net/humanize.
Read on 15 September 2026: Turnitin's AI writing detection capabilities FAQ (stamped Updated 17 days ago), gptzero.me/technology and gptzero.me/faq. All three in a real browser, on rendered text, with a positive and a negative control on every count.
Top comments (0)